Multivariate Time Series Classification - Activity Data from Wireless Sensor Network
- This project tackles human-activity recognition on the UCI AReM wireless-sensor dataset using a classical multivariate time-series pipeline. Time-domain features are engineered per channel and within-instance dynamics are captured by segmenting each series into ℓ chunks. Models compare (i) binary bending vs. non-bending using p-value-guided RFE + logistic regression vs. L1-penalized LR, and (ii) multiclass activity recognition using multinomial LR (L1) and Naive Bayes, with a Gaussian NB + PCA variant to decorrelate features. Selection and evaluation use nested/outer CV with ROC/AUC, accuracy, and confusion matrices.
- Adding PCA to a Gaussian NB pipeline improved multiclass CV test accuracy to .84
Presentation | Github